| 摘要: |
| 【目的】提出了一种改进的YOLOv4模型，为自然环境下3种常见茶叶病害（茶白星病、茶云纹叶枯病和茶轮斑病）的快速精准识别提供支持。【方法】使用MobileNetv2和深度可分离卷积来降低YOLOv4模型的参数量，并引入卷积注意力模块对YOLOv4模型进行识别精度改进。采用平均精度、平均精度均值、图像检测速度和模型大小作为模型性能评价指标，在相同的茶叶病害数据集和试验平台中，对改进YOLOv4模型与原始YOLOv4模型、其他目标检测模型（YOLOv3、SSD和Faster R CNN）的病害识别效果进行对比试验。【结果】与原始YOLOv4模型相比，改进YOLOv4模型的大小减少了83.2%，对茶白星病、茶云纹叶枯病和茶轮斑病识别的平均精度分别提高了6.2%，1.7%和1.6%，平均精度均值达到93.85%，图像检测速度为26.6帧/s。与YOLOv3、SSD和Faster R-CNN模型相比，改进YOLOv4模型的平均精度均值分别提高了6.0%，13.7%和3.4%，图像检测速度分别提高了5.5，7.3和11.7帧/s。【结论】对YOLOv4模型所使用的改进方法具备有效性，所提出的改进YOLOv4模型可以实现对自然环境下3种常见茶叶病害的快速精准识别。 |
| 关键词: 茶白星病 茶云纹叶枯病 茶轮斑病 YOLOv4模型 茶叶病害识别 |
| DOI： |
| 分类号: |
| 基金项目:广东省现代农业关键技术模式集成与示范推广项目（粤财农［2021］37号-200011）；国家自然科学基金项目（31671591，31971797）；广州市科技计划项目（202002030245）；广东省科技专项资金项目（“大专项+任务清单”）（2020020103）；广东省现代农业产业技术体系创新团队建设专项资金项目（2022KJ108）；广东省教育厅特色创新类项目（2019KTSCX013）；2020年广东省科技创新战略专项资金项目（“攀登计划”，pdjh2020a0084）；广东省大学生创新创业项目（S202010564150，202110564042） |
|
| Recognition of tea diseases based on improved YOLOv4 model |
|
SUN Daozong,LIU Huan,LIU Jinyuan,et al
|
| Abstract: |
| 【Objective】An improved YOLOv4 model was proposed to provide support for the rapid and accurate recognition of three common tea diseases of tea white scab disease,tea cloud leaf blight and tea ring spot in natural environment.【Method】The number of parameters of the YOLOv4 model was reduced by using MobileNetv2 and depthwise separable convolution,and the convolutional block attention module was introduced to improve recognition precision of the model.Using average precision,mean average precision,image detection speed and model size as evaluation indexes,the disease recognition ability of the improved YOLOv4 model was compared with that of the original YOLOv4 model and other target detection models (YOLOv3,SSD and Faster R-CNN) using the same tea disease dataset and testbed.【Result】Compared with the original YOLOv4 model,the size of the improved YOLOv4 model was decreased by 83.2%, and the average recognition precision of tea white scab disease,tea cloud leaf blight and tea ring spot was increased by 6.2%,1.7% and 1.6%,respectively.The mean average precision reached 93.85%,and the image detection speed was 26.6 frames/s.Compared with YOLOv3,SSD and Faster R-CNN models,the mean average precision of the improved YOLOv4 model was increased by 6.0%,13.7% and 3.4%,and the image detection speed was increased by 5.5,7.3 and 11.7 frames/s,respectively.【Conclusion】The improved methods were valid,and the improved YOLOv4 model achieved rapid and accurate recognition of three common tea diseases in natural environment. |
| Key words: tea white scab disease tea cloud leaf blight tea ring spot YOLOv4 model recognition of tea diseases |
